1 /****************************************************************************
2 * Driver for Solarflare Solarstorm network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2011 Solarflare Communications Inc.
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation, incorporated herein by reference.
11 #include <linux/module.h>
12 #include <linux/pci.h>
13 #include <linux/netdevice.h>
14 #include <linux/etherdevice.h>
15 #include <linux/delay.h>
16 #include <linux/notifier.h>
18 #include <linux/tcp.h>
20 #include <linux/crc32.h>
21 #include <linux/ethtool.h>
22 #include <linux/topology.h>
23 #include <linux/gfp.h>
24 #include <linux/cpu_rmap.h>
25 #include "net_driver.h"
30 #include "workarounds.h"
32 /**************************************************************************
36 **************************************************************************
39 /* Loopback mode names (see LOOPBACK_MODE()) */
40 const unsigned int efx_loopback_mode_max
= LOOPBACK_MAX
;
41 const char *const efx_loopback_mode_names
[] = {
42 [LOOPBACK_NONE
] = "NONE",
43 [LOOPBACK_DATA
] = "DATAPATH",
44 [LOOPBACK_GMAC
] = "GMAC",
45 [LOOPBACK_XGMII
] = "XGMII",
46 [LOOPBACK_XGXS
] = "XGXS",
47 [LOOPBACK_XAUI
] = "XAUI",
48 [LOOPBACK_GMII
] = "GMII",
49 [LOOPBACK_SGMII
] = "SGMII",
50 [LOOPBACK_XGBR
] = "XGBR",
51 [LOOPBACK_XFI
] = "XFI",
52 [LOOPBACK_XAUI_FAR
] = "XAUI_FAR",
53 [LOOPBACK_GMII_FAR
] = "GMII_FAR",
54 [LOOPBACK_SGMII_FAR
] = "SGMII_FAR",
55 [LOOPBACK_XFI_FAR
] = "XFI_FAR",
56 [LOOPBACK_GPHY
] = "GPHY",
57 [LOOPBACK_PHYXS
] = "PHYXS",
58 [LOOPBACK_PCS
] = "PCS",
59 [LOOPBACK_PMAPMD
] = "PMA/PMD",
60 [LOOPBACK_XPORT
] = "XPORT",
61 [LOOPBACK_XGMII_WS
] = "XGMII_WS",
62 [LOOPBACK_XAUI_WS
] = "XAUI_WS",
63 [LOOPBACK_XAUI_WS_FAR
] = "XAUI_WS_FAR",
64 [LOOPBACK_XAUI_WS_NEAR
] = "XAUI_WS_NEAR",
65 [LOOPBACK_GMII_WS
] = "GMII_WS",
66 [LOOPBACK_XFI_WS
] = "XFI_WS",
67 [LOOPBACK_XFI_WS_FAR
] = "XFI_WS_FAR",
68 [LOOPBACK_PHYXS_WS
] = "PHYXS_WS",
71 const unsigned int efx_reset_type_max
= RESET_TYPE_MAX
;
72 const char *const efx_reset_type_names
[] = {
73 [RESET_TYPE_INVISIBLE
] = "INVISIBLE",
74 [RESET_TYPE_ALL
] = "ALL",
75 [RESET_TYPE_WORLD
] = "WORLD",
76 [RESET_TYPE_DISABLE
] = "DISABLE",
77 [RESET_TYPE_TX_WATCHDOG
] = "TX_WATCHDOG",
78 [RESET_TYPE_INT_ERROR
] = "INT_ERROR",
79 [RESET_TYPE_RX_RECOVERY
] = "RX_RECOVERY",
80 [RESET_TYPE_RX_DESC_FETCH
] = "RX_DESC_FETCH",
81 [RESET_TYPE_TX_DESC_FETCH
] = "TX_DESC_FETCH",
82 [RESET_TYPE_TX_SKIP
] = "TX_SKIP",
83 [RESET_TYPE_MC_FAILURE
] = "MC_FAILURE",
86 #define EFX_MAX_MTU (9 * 1024)
88 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
89 * queued onto this work queue. This is not a per-nic work queue, because
90 * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
92 static struct workqueue_struct
*reset_workqueue
;
94 /**************************************************************************
98 *************************************************************************/
101 * Use separate channels for TX and RX events
103 * Set this to 1 to use separate channels for TX and RX. It allows us
104 * to control interrupt affinity separately for TX and RX.
106 * This is only used in MSI-X interrupt mode
108 static unsigned int separate_tx_channels
;
109 module_param(separate_tx_channels
, uint
, 0444);
110 MODULE_PARM_DESC(separate_tx_channels
,
111 "Use separate channels for TX and RX");
113 /* This is the weight assigned to each of the (per-channel) virtual
116 static int napi_weight
= 64;
118 /* This is the time (in jiffies) between invocations of the hardware
119 * monitor. On Falcon-based NICs, this will:
120 * - Check the on-board hardware monitor;
121 * - Poll the link state and reconfigure the hardware as necessary.
123 static unsigned int efx_monitor_interval
= 1 * HZ
;
125 /* Initial interrupt moderation settings. They can be modified after
126 * module load with ethtool.
128 * The default for RX should strike a balance between increasing the
129 * round-trip latency and reducing overhead.
131 static unsigned int rx_irq_mod_usec
= 60;
133 /* Initial interrupt moderation settings. They can be modified after
134 * module load with ethtool.
136 * This default is chosen to ensure that a 10G link does not go idle
137 * while a TX queue is stopped after it has become full. A queue is
138 * restarted when it drops below half full. The time this takes (assuming
139 * worst case 3 descriptors per packet and 1024 descriptors) is
140 * 512 / 3 * 1.2 = 205 usec.
142 static unsigned int tx_irq_mod_usec
= 150;
144 /* This is the first interrupt mode to try out of:
149 static unsigned int interrupt_mode
;
151 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
152 * i.e. the number of CPUs among which we may distribute simultaneous
153 * interrupt handling.
155 * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
156 * The default (0) means to assign an interrupt to each core.
158 static unsigned int rss_cpus
;
159 module_param(rss_cpus
, uint
, 0444);
160 MODULE_PARM_DESC(rss_cpus
, "Number of CPUs to use for Receive-Side Scaling");
162 static int phy_flash_cfg
;
163 module_param(phy_flash_cfg
, int, 0644);
164 MODULE_PARM_DESC(phy_flash_cfg
, "Set PHYs into reflash mode initially");
166 static unsigned irq_adapt_low_thresh
= 10000;
167 module_param(irq_adapt_low_thresh
, uint
, 0644);
168 MODULE_PARM_DESC(irq_adapt_low_thresh
,
169 "Threshold score for reducing IRQ moderation");
171 static unsigned irq_adapt_high_thresh
= 20000;
172 module_param(irq_adapt_high_thresh
, uint
, 0644);
173 MODULE_PARM_DESC(irq_adapt_high_thresh
,
174 "Threshold score for increasing IRQ moderation");
176 static unsigned debug
= (NETIF_MSG_DRV
| NETIF_MSG_PROBE
|
177 NETIF_MSG_LINK
| NETIF_MSG_IFDOWN
|
178 NETIF_MSG_IFUP
| NETIF_MSG_RX_ERR
|
179 NETIF_MSG_TX_ERR
| NETIF_MSG_HW
);
180 module_param(debug
, uint
, 0);
181 MODULE_PARM_DESC(debug
, "Bitmapped debugging message enable value");
183 /**************************************************************************
185 * Utility functions and prototypes
187 *************************************************************************/
189 static void efx_start_interrupts(struct efx_nic
*efx
);
190 static void efx_stop_interrupts(struct efx_nic
*efx
);
191 static void efx_remove_channels(struct efx_nic
*efx
);
192 static void efx_remove_port(struct efx_nic
*efx
);
193 static void efx_init_napi(struct efx_nic
*efx
);
194 static void efx_fini_napi(struct efx_nic
*efx
);
195 static void efx_fini_napi_channel(struct efx_channel
*channel
);
196 static void efx_fini_struct(struct efx_nic
*efx
);
197 static void efx_start_all(struct efx_nic
*efx
);
198 static void efx_stop_all(struct efx_nic
*efx
);
200 #define EFX_ASSERT_RESET_SERIALISED(efx) \
202 if ((efx->state == STATE_RUNNING) || \
203 (efx->state == STATE_DISABLED)) \
207 /**************************************************************************
209 * Event queue processing
211 *************************************************************************/
213 /* Process channel's event queue
215 * This function is responsible for processing the event queue of a
216 * single channel. The caller must guarantee that this function will
217 * never be concurrently called more than once on the same channel,
218 * though different channels may be being processed concurrently.
220 static int efx_process_channel(struct efx_channel
*channel
, int budget
)
224 if (unlikely(!channel
->enabled
))
227 spent
= efx_nic_process_eventq(channel
, budget
);
228 if (spent
&& efx_channel_has_rx_queue(channel
)) {
229 struct efx_rx_queue
*rx_queue
=
230 efx_channel_get_rx_queue(channel
);
232 /* Deliver last RX packet. */
233 if (channel
->rx_pkt
) {
234 __efx_rx_packet(channel
, channel
->rx_pkt
);
235 channel
->rx_pkt
= NULL
;
237 if (rx_queue
->enabled
) {
238 efx_rx_strategy(channel
);
239 efx_fast_push_rx_descriptors(rx_queue
);
246 /* Mark channel as finished processing
248 * Note that since we will not receive further interrupts for this
249 * channel before we finish processing and call the eventq_read_ack()
250 * method, there is no need to use the interrupt hold-off timers.
252 static inline void efx_channel_processed(struct efx_channel
*channel
)
254 /* The interrupt handler for this channel may set work_pending
255 * as soon as we acknowledge the events we've seen. Make sure
256 * it's cleared before then. */
257 channel
->work_pending
= false;
260 efx_nic_eventq_read_ack(channel
);
265 * NAPI guarantees serialisation of polls of the same device, which
266 * provides the guarantee required by efx_process_channel().
268 static int efx_poll(struct napi_struct
*napi
, int budget
)
270 struct efx_channel
*channel
=
271 container_of(napi
, struct efx_channel
, napi_str
);
272 struct efx_nic
*efx
= channel
->efx
;
275 netif_vdbg(efx
, intr
, efx
->net_dev
,
276 "channel %d NAPI poll executing on CPU %d\n",
277 channel
->channel
, raw_smp_processor_id());
279 spent
= efx_process_channel(channel
, budget
);
281 if (spent
< budget
) {
282 if (channel
->channel
< efx
->n_rx_channels
&&
283 efx
->irq_rx_adaptive
&&
284 unlikely(++channel
->irq_count
== 1000)) {
285 if (unlikely(channel
->irq_mod_score
<
286 irq_adapt_low_thresh
)) {
287 if (channel
->irq_moderation
> 1) {
288 channel
->irq_moderation
-= 1;
289 efx
->type
->push_irq_moderation(channel
);
291 } else if (unlikely(channel
->irq_mod_score
>
292 irq_adapt_high_thresh
)) {
293 if (channel
->irq_moderation
<
294 efx
->irq_rx_moderation
) {
295 channel
->irq_moderation
+= 1;
296 efx
->type
->push_irq_moderation(channel
);
299 channel
->irq_count
= 0;
300 channel
->irq_mod_score
= 0;
303 efx_filter_rfs_expire(channel
);
305 /* There is no race here; although napi_disable() will
306 * only wait for napi_complete(), this isn't a problem
307 * since efx_channel_processed() will have no effect if
308 * interrupts have already been disabled.
311 efx_channel_processed(channel
);
317 /* Process the eventq of the specified channel immediately on this CPU
319 * Disable hardware generated interrupts, wait for any existing
320 * processing to finish, then directly poll (and ack ) the eventq.
321 * Finally reenable NAPI and interrupts.
323 * This is for use only during a loopback self-test. It must not
324 * deliver any packets up the stack as this can result in deadlock.
326 void efx_process_channel_now(struct efx_channel
*channel
)
328 struct efx_nic
*efx
= channel
->efx
;
330 BUG_ON(channel
->channel
>= efx
->n_channels
);
331 BUG_ON(!channel
->enabled
);
332 BUG_ON(!efx
->loopback_selftest
);
334 /* Disable interrupts and wait for ISRs to complete */
335 efx_nic_disable_interrupts(efx
);
336 if (efx
->legacy_irq
) {
337 synchronize_irq(efx
->legacy_irq
);
338 efx
->legacy_irq_enabled
= false;
341 synchronize_irq(channel
->irq
);
343 /* Wait for any NAPI processing to complete */
344 napi_disable(&channel
->napi_str
);
346 /* Poll the channel */
347 efx_process_channel(channel
, channel
->eventq_mask
+ 1);
349 /* Ack the eventq. This may cause an interrupt to be generated
350 * when they are reenabled */
351 efx_channel_processed(channel
);
353 napi_enable(&channel
->napi_str
);
355 efx
->legacy_irq_enabled
= true;
356 efx_nic_enable_interrupts(efx
);
359 /* Create event queue
360 * Event queue memory allocations are done only once. If the channel
361 * is reset, the memory buffer will be reused; this guards against
362 * errors during channel reset and also simplifies interrupt handling.
364 static int efx_probe_eventq(struct efx_channel
*channel
)
366 struct efx_nic
*efx
= channel
->efx
;
367 unsigned long entries
;
369 netif_dbg(efx
, probe
, efx
->net_dev
,
370 "chan %d create event queue\n", channel
->channel
);
372 /* Build an event queue with room for one event per tx and rx buffer,
373 * plus some extra for link state events and MCDI completions. */
374 entries
= roundup_pow_of_two(efx
->rxq_entries
+ efx
->txq_entries
+ 128);
375 EFX_BUG_ON_PARANOID(entries
> EFX_MAX_EVQ_SIZE
);
376 channel
->eventq_mask
= max(entries
, EFX_MIN_EVQ_SIZE
) - 1;
378 return efx_nic_probe_eventq(channel
);
381 /* Prepare channel's event queue */
382 static void efx_init_eventq(struct efx_channel
*channel
)
384 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
385 "chan %d init event queue\n", channel
->channel
);
387 channel
->eventq_read_ptr
= 0;
389 efx_nic_init_eventq(channel
);
392 /* Enable event queue processing and NAPI */
393 static void efx_start_eventq(struct efx_channel
*channel
)
395 netif_dbg(channel
->efx
, ifup
, channel
->efx
->net_dev
,
396 "chan %d start event queue\n", channel
->channel
);
398 /* The interrupt handler for this channel may set work_pending
399 * as soon as we enable it. Make sure it's cleared before
400 * then. Similarly, make sure it sees the enabled flag set.
402 channel
->work_pending
= false;
403 channel
->enabled
= true;
406 napi_enable(&channel
->napi_str
);
407 efx_nic_eventq_read_ack(channel
);
410 /* Disable event queue processing and NAPI */
411 static void efx_stop_eventq(struct efx_channel
*channel
)
413 if (!channel
->enabled
)
416 napi_disable(&channel
->napi_str
);
417 channel
->enabled
= false;
420 static void efx_fini_eventq(struct efx_channel
*channel
)
422 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
423 "chan %d fini event queue\n", channel
->channel
);
425 efx_nic_fini_eventq(channel
);
428 static void efx_remove_eventq(struct efx_channel
*channel
)
430 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
431 "chan %d remove event queue\n", channel
->channel
);
433 efx_nic_remove_eventq(channel
);
436 /**************************************************************************
440 *************************************************************************/
442 /* Allocate and initialise a channel structure, optionally copying
443 * parameters (but not resources) from an old channel structure. */
444 static struct efx_channel
*
445 efx_alloc_channel(struct efx_nic
*efx
, int i
, struct efx_channel
*old_channel
)
447 struct efx_channel
*channel
;
448 struct efx_rx_queue
*rx_queue
;
449 struct efx_tx_queue
*tx_queue
;
453 channel
= kmalloc(sizeof(*channel
), GFP_KERNEL
);
457 *channel
= *old_channel
;
459 channel
->napi_dev
= NULL
;
460 memset(&channel
->eventq
, 0, sizeof(channel
->eventq
));
462 rx_queue
= &channel
->rx_queue
;
463 rx_queue
->buffer
= NULL
;
464 memset(&rx_queue
->rxd
, 0, sizeof(rx_queue
->rxd
));
466 for (j
= 0; j
< EFX_TXQ_TYPES
; j
++) {
467 tx_queue
= &channel
->tx_queue
[j
];
468 if (tx_queue
->channel
)
469 tx_queue
->channel
= channel
;
470 tx_queue
->buffer
= NULL
;
471 memset(&tx_queue
->txd
, 0, sizeof(tx_queue
->txd
));
474 channel
= kzalloc(sizeof(*channel
), GFP_KERNEL
);
479 channel
->channel
= i
;
481 for (j
= 0; j
< EFX_TXQ_TYPES
; j
++) {
482 tx_queue
= &channel
->tx_queue
[j
];
484 tx_queue
->queue
= i
* EFX_TXQ_TYPES
+ j
;
485 tx_queue
->channel
= channel
;
489 rx_queue
= &channel
->rx_queue
;
491 setup_timer(&rx_queue
->slow_fill
, efx_rx_slow_fill
,
492 (unsigned long)rx_queue
);
497 static int efx_probe_channel(struct efx_channel
*channel
)
499 struct efx_tx_queue
*tx_queue
;
500 struct efx_rx_queue
*rx_queue
;
503 netif_dbg(channel
->efx
, probe
, channel
->efx
->net_dev
,
504 "creating channel %d\n", channel
->channel
);
506 rc
= efx_probe_eventq(channel
);
510 efx_for_each_channel_tx_queue(tx_queue
, channel
) {
511 rc
= efx_probe_tx_queue(tx_queue
);
516 efx_for_each_channel_rx_queue(rx_queue
, channel
) {
517 rc
= efx_probe_rx_queue(rx_queue
);
522 channel
->n_rx_frm_trunc
= 0;
527 efx_for_each_channel_rx_queue(rx_queue
, channel
)
528 efx_remove_rx_queue(rx_queue
);
530 efx_for_each_channel_tx_queue(tx_queue
, channel
)
531 efx_remove_tx_queue(tx_queue
);
537 static void efx_set_channel_names(struct efx_nic
*efx
)
539 struct efx_channel
*channel
;
540 const char *type
= "";
543 efx_for_each_channel(channel
, efx
) {
544 number
= channel
->channel
;
545 if (efx
->n_channels
> efx
->n_rx_channels
) {
546 if (channel
->channel
< efx
->n_rx_channels
) {
550 number
-= efx
->n_rx_channels
;
553 snprintf(efx
->channel_name
[channel
->channel
],
554 sizeof(efx
->channel_name
[0]),
555 "%s%s-%d", efx
->name
, type
, number
);
559 static int efx_probe_channels(struct efx_nic
*efx
)
561 struct efx_channel
*channel
;
564 /* Restart special buffer allocation */
565 efx
->next_buffer_table
= 0;
567 efx_for_each_channel(channel
, efx
) {
568 rc
= efx_probe_channel(channel
);
570 netif_err(efx
, probe
, efx
->net_dev
,
571 "failed to create channel %d\n",
576 efx_set_channel_names(efx
);
581 efx_remove_channels(efx
);
585 /* Channels are shutdown and reinitialised whilst the NIC is running
586 * to propagate configuration changes (mtu, checksum offload), or
587 * to clear hardware error conditions
589 static void efx_start_datapath(struct efx_nic
*efx
)
591 struct efx_tx_queue
*tx_queue
;
592 struct efx_rx_queue
*rx_queue
;
593 struct efx_channel
*channel
;
595 /* Calculate the rx buffer allocation parameters required to
596 * support the current MTU, including padding for header
597 * alignment and overruns.
599 efx
->rx_buffer_len
= (max(EFX_PAGE_IP_ALIGN
, NET_IP_ALIGN
) +
600 EFX_MAX_FRAME_LEN(efx
->net_dev
->mtu
) +
601 efx
->type
->rx_buffer_hash_size
+
602 efx
->type
->rx_buffer_padding
);
603 efx
->rx_buffer_order
= get_order(efx
->rx_buffer_len
+
604 sizeof(struct efx_rx_page_state
));
606 /* Initialise the channels */
607 efx_for_each_channel(channel
, efx
) {
608 efx_for_each_channel_tx_queue(tx_queue
, channel
)
609 efx_init_tx_queue(tx_queue
);
611 /* The rx buffer allocation strategy is MTU dependent */
612 efx_rx_strategy(channel
);
614 efx_for_each_channel_rx_queue(rx_queue
, channel
) {
615 efx_init_rx_queue(rx_queue
);
616 efx_nic_generate_fill_event(rx_queue
);
619 WARN_ON(channel
->rx_pkt
!= NULL
);
620 efx_rx_strategy(channel
);
623 if (netif_device_present(efx
->net_dev
))
624 netif_tx_wake_all_queues(efx
->net_dev
);
627 static void efx_stop_datapath(struct efx_nic
*efx
)
629 struct efx_channel
*channel
;
630 struct efx_tx_queue
*tx_queue
;
631 struct efx_rx_queue
*rx_queue
;
634 EFX_ASSERT_RESET_SERIALISED(efx
);
635 BUG_ON(efx
->port_enabled
);
637 rc
= efx_nic_flush_queues(efx
);
638 if (rc
&& EFX_WORKAROUND_7803(efx
)) {
639 /* Schedule a reset to recover from the flush failure. The
640 * descriptor caches reference memory we're about to free,
641 * but falcon_reconfigure_mac_wrapper() won't reconnect
642 * the MACs because of the pending reset. */
643 netif_err(efx
, drv
, efx
->net_dev
,
644 "Resetting to recover from flush failure\n");
645 efx_schedule_reset(efx
, RESET_TYPE_ALL
);
647 netif_err(efx
, drv
, efx
->net_dev
, "failed to flush queues\n");
649 netif_dbg(efx
, drv
, efx
->net_dev
,
650 "successfully flushed all queues\n");
653 efx_for_each_channel(channel
, efx
) {
654 /* RX packet processing is pipelined, so wait for the
655 * NAPI handler to complete. At least event queue 0
656 * might be kept active by non-data events, so don't
657 * use napi_synchronize() but actually disable NAPI
660 if (efx_channel_has_rx_queue(channel
)) {
661 efx_stop_eventq(channel
);
662 efx_start_eventq(channel
);
665 efx_for_each_channel_rx_queue(rx_queue
, channel
)
666 efx_fini_rx_queue(rx_queue
);
667 efx_for_each_possible_channel_tx_queue(tx_queue
, channel
)
668 efx_fini_tx_queue(tx_queue
);
672 static void efx_remove_channel(struct efx_channel
*channel
)
674 struct efx_tx_queue
*tx_queue
;
675 struct efx_rx_queue
*rx_queue
;
677 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
678 "destroy chan %d\n", channel
->channel
);
680 efx_for_each_channel_rx_queue(rx_queue
, channel
)
681 efx_remove_rx_queue(rx_queue
);
682 efx_for_each_possible_channel_tx_queue(tx_queue
, channel
)
683 efx_remove_tx_queue(tx_queue
);
684 efx_remove_eventq(channel
);
687 static void efx_remove_channels(struct efx_nic
*efx
)
689 struct efx_channel
*channel
;
691 efx_for_each_channel(channel
, efx
)
692 efx_remove_channel(channel
);
696 efx_realloc_channels(struct efx_nic
*efx
, u32 rxq_entries
, u32 txq_entries
)
698 struct efx_channel
*other_channel
[EFX_MAX_CHANNELS
], *channel
;
699 u32 old_rxq_entries
, old_txq_entries
;
704 efx_stop_interrupts(efx
);
707 memset(other_channel
, 0, sizeof(other_channel
));
708 for (i
= 0; i
< efx
->n_channels
; i
++) {
709 channel
= efx_alloc_channel(efx
, i
, efx
->channel
[i
]);
714 other_channel
[i
] = channel
;
717 /* Swap entry counts and channel pointers */
718 old_rxq_entries
= efx
->rxq_entries
;
719 old_txq_entries
= efx
->txq_entries
;
720 efx
->rxq_entries
= rxq_entries
;
721 efx
->txq_entries
= txq_entries
;
722 for (i
= 0; i
< efx
->n_channels
; i
++) {
723 channel
= efx
->channel
[i
];
724 efx
->channel
[i
] = other_channel
[i
];
725 other_channel
[i
] = channel
;
728 rc
= efx_probe_channels(efx
);
734 /* Destroy old channels */
735 for (i
= 0; i
< efx
->n_channels
; i
++) {
736 efx_fini_napi_channel(other_channel
[i
]);
737 efx_remove_channel(other_channel
[i
]);
740 /* Free unused channel structures */
741 for (i
= 0; i
< efx
->n_channels
; i
++)
742 kfree(other_channel
[i
]);
744 efx_start_interrupts(efx
);
750 efx
->rxq_entries
= old_rxq_entries
;
751 efx
->txq_entries
= old_txq_entries
;
752 for (i
= 0; i
< efx
->n_channels
; i
++) {
753 channel
= efx
->channel
[i
];
754 efx
->channel
[i
] = other_channel
[i
];
755 other_channel
[i
] = channel
;
760 void efx_schedule_slow_fill(struct efx_rx_queue
*rx_queue
)
762 mod_timer(&rx_queue
->slow_fill
, jiffies
+ msecs_to_jiffies(100));
765 /**************************************************************************
769 **************************************************************************/
771 /* This ensures that the kernel is kept informed (via
772 * netif_carrier_on/off) of the link status, and also maintains the
773 * link status's stop on the port's TX queue.
775 void efx_link_status_changed(struct efx_nic
*efx
)
777 struct efx_link_state
*link_state
= &efx
->link_state
;
779 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
780 * that no events are triggered between unregister_netdev() and the
781 * driver unloading. A more general condition is that NETDEV_CHANGE
782 * can only be generated between NETDEV_UP and NETDEV_DOWN */
783 if (!netif_running(efx
->net_dev
))
786 if (link_state
->up
!= netif_carrier_ok(efx
->net_dev
)) {
787 efx
->n_link_state_changes
++;
790 netif_carrier_on(efx
->net_dev
);
792 netif_carrier_off(efx
->net_dev
);
795 /* Status message for kernel log */
797 netif_info(efx
, link
, efx
->net_dev
,
798 "link up at %uMbps %s-duplex (MTU %d)%s\n",
799 link_state
->speed
, link_state
->fd
? "full" : "half",
801 (efx
->promiscuous
? " [PROMISC]" : ""));
803 netif_info(efx
, link
, efx
->net_dev
, "link down\n");
806 void efx_link_set_advertising(struct efx_nic
*efx
, u32 advertising
)
808 efx
->link_advertising
= advertising
;
810 if (advertising
& ADVERTISED_Pause
)
811 efx
->wanted_fc
|= (EFX_FC_TX
| EFX_FC_RX
);
813 efx
->wanted_fc
&= ~(EFX_FC_TX
| EFX_FC_RX
);
814 if (advertising
& ADVERTISED_Asym_Pause
)
815 efx
->wanted_fc
^= EFX_FC_TX
;
819 void efx_link_set_wanted_fc(struct efx_nic
*efx
, u8 wanted_fc
)
821 efx
->wanted_fc
= wanted_fc
;
822 if (efx
->link_advertising
) {
823 if (wanted_fc
& EFX_FC_RX
)
824 efx
->link_advertising
|= (ADVERTISED_Pause
|
825 ADVERTISED_Asym_Pause
);
827 efx
->link_advertising
&= ~(ADVERTISED_Pause
|
828 ADVERTISED_Asym_Pause
);
829 if (wanted_fc
& EFX_FC_TX
)
830 efx
->link_advertising
^= ADVERTISED_Asym_Pause
;
834 static void efx_fini_port(struct efx_nic
*efx
);
836 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
837 * the MAC appropriately. All other PHY configuration changes are pushed
838 * through phy_op->set_settings(), and pushed asynchronously to the MAC
839 * through efx_monitor().
841 * Callers must hold the mac_lock
843 int __efx_reconfigure_port(struct efx_nic
*efx
)
845 enum efx_phy_mode phy_mode
;
848 WARN_ON(!mutex_is_locked(&efx
->mac_lock
));
850 /* Serialise the promiscuous flag with efx_set_rx_mode. */
851 netif_addr_lock_bh(efx
->net_dev
);
852 netif_addr_unlock_bh(efx
->net_dev
);
854 /* Disable PHY transmit in mac level loopbacks */
855 phy_mode
= efx
->phy_mode
;
856 if (LOOPBACK_INTERNAL(efx
))
857 efx
->phy_mode
|= PHY_MODE_TX_DISABLED
;
859 efx
->phy_mode
&= ~PHY_MODE_TX_DISABLED
;
861 rc
= efx
->type
->reconfigure_port(efx
);
864 efx
->phy_mode
= phy_mode
;
869 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
871 int efx_reconfigure_port(struct efx_nic
*efx
)
875 EFX_ASSERT_RESET_SERIALISED(efx
);
877 mutex_lock(&efx
->mac_lock
);
878 rc
= __efx_reconfigure_port(efx
);
879 mutex_unlock(&efx
->mac_lock
);
884 /* Asynchronous work item for changing MAC promiscuity and multicast
885 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
887 static void efx_mac_work(struct work_struct
*data
)
889 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, mac_work
);
891 mutex_lock(&efx
->mac_lock
);
892 if (efx
->port_enabled
)
893 efx
->type
->reconfigure_mac(efx
);
894 mutex_unlock(&efx
->mac_lock
);
897 static int efx_probe_port(struct efx_nic
*efx
)
901 netif_dbg(efx
, probe
, efx
->net_dev
, "create port\n");
904 efx
->phy_mode
= PHY_MODE_SPECIAL
;
906 /* Connect up MAC/PHY operations table */
907 rc
= efx
->type
->probe_port(efx
);
911 /* Initialise MAC address to permanent address */
912 memcpy(efx
->net_dev
->dev_addr
, efx
->net_dev
->perm_addr
, ETH_ALEN
);
917 static int efx_init_port(struct efx_nic
*efx
)
921 netif_dbg(efx
, drv
, efx
->net_dev
, "init port\n");
923 mutex_lock(&efx
->mac_lock
);
925 rc
= efx
->phy_op
->init(efx
);
929 efx
->port_initialized
= true;
931 /* Reconfigure the MAC before creating dma queues (required for
932 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
933 efx
->type
->reconfigure_mac(efx
);
935 /* Ensure the PHY advertises the correct flow control settings */
936 rc
= efx
->phy_op
->reconfigure(efx
);
940 mutex_unlock(&efx
->mac_lock
);
944 efx
->phy_op
->fini(efx
);
946 mutex_unlock(&efx
->mac_lock
);
950 static void efx_start_port(struct efx_nic
*efx
)
952 netif_dbg(efx
, ifup
, efx
->net_dev
, "start port\n");
953 BUG_ON(efx
->port_enabled
);
955 mutex_lock(&efx
->mac_lock
);
956 efx
->port_enabled
= true;
958 /* efx_mac_work() might have been scheduled after efx_stop_port(),
959 * and then cancelled by efx_flush_all() */
960 efx
->type
->reconfigure_mac(efx
);
962 mutex_unlock(&efx
->mac_lock
);
965 /* Prevent efx_mac_work() and efx_monitor() from working */
966 static void efx_stop_port(struct efx_nic
*efx
)
968 netif_dbg(efx
, ifdown
, efx
->net_dev
, "stop port\n");
970 mutex_lock(&efx
->mac_lock
);
971 efx
->port_enabled
= false;
972 mutex_unlock(&efx
->mac_lock
);
974 /* Serialise against efx_set_multicast_list() */
975 netif_addr_lock_bh(efx
->net_dev
);
976 netif_addr_unlock_bh(efx
->net_dev
);
979 static void efx_fini_port(struct efx_nic
*efx
)
981 netif_dbg(efx
, drv
, efx
->net_dev
, "shut down port\n");
983 if (!efx
->port_initialized
)
986 efx
->phy_op
->fini(efx
);
987 efx
->port_initialized
= false;
989 efx
->link_state
.up
= false;
990 efx_link_status_changed(efx
);
993 static void efx_remove_port(struct efx_nic
*efx
)
995 netif_dbg(efx
, drv
, efx
->net_dev
, "destroying port\n");
997 efx
->type
->remove_port(efx
);
1000 /**************************************************************************
1004 **************************************************************************/
1006 /* This configures the PCI device to enable I/O and DMA. */
1007 static int efx_init_io(struct efx_nic
*efx
)
1009 struct pci_dev
*pci_dev
= efx
->pci_dev
;
1010 dma_addr_t dma_mask
= efx
->type
->max_dma_mask
;
1013 netif_dbg(efx
, probe
, efx
->net_dev
, "initialising I/O\n");
1015 rc
= pci_enable_device(pci_dev
);
1017 netif_err(efx
, probe
, efx
->net_dev
,
1018 "failed to enable PCI device\n");
1022 pci_set_master(pci_dev
);
1024 /* Set the PCI DMA mask. Try all possibilities from our
1025 * genuine mask down to 32 bits, because some architectures
1026 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1027 * masks event though they reject 46 bit masks.
1029 while (dma_mask
> 0x7fffffffUL
) {
1030 if (pci_dma_supported(pci_dev
, dma_mask
)) {
1031 rc
= pci_set_dma_mask(pci_dev
, dma_mask
);
1038 netif_err(efx
, probe
, efx
->net_dev
,
1039 "could not find a suitable DMA mask\n");
1042 netif_dbg(efx
, probe
, efx
->net_dev
,
1043 "using DMA mask %llx\n", (unsigned long long) dma_mask
);
1044 rc
= pci_set_consistent_dma_mask(pci_dev
, dma_mask
);
1046 /* pci_set_consistent_dma_mask() is not *allowed* to
1047 * fail with a mask that pci_set_dma_mask() accepted,
1048 * but just in case...
1050 netif_err(efx
, probe
, efx
->net_dev
,
1051 "failed to set consistent DMA mask\n");
1055 efx
->membase_phys
= pci_resource_start(efx
->pci_dev
, EFX_MEM_BAR
);
1056 rc
= pci_request_region(pci_dev
, EFX_MEM_BAR
, "sfc");
1058 netif_err(efx
, probe
, efx
->net_dev
,
1059 "request for memory BAR failed\n");
1063 efx
->membase
= ioremap_nocache(efx
->membase_phys
,
1064 efx
->type
->mem_map_size
);
1065 if (!efx
->membase
) {
1066 netif_err(efx
, probe
, efx
->net_dev
,
1067 "could not map memory BAR at %llx+%x\n",
1068 (unsigned long long)efx
->membase_phys
,
1069 efx
->type
->mem_map_size
);
1073 netif_dbg(efx
, probe
, efx
->net_dev
,
1074 "memory BAR at %llx+%x (virtual %p)\n",
1075 (unsigned long long)efx
->membase_phys
,
1076 efx
->type
->mem_map_size
, efx
->membase
);
1081 pci_release_region(efx
->pci_dev
, EFX_MEM_BAR
);
1083 efx
->membase_phys
= 0;
1085 pci_disable_device(efx
->pci_dev
);
1090 static void efx_fini_io(struct efx_nic
*efx
)
1092 netif_dbg(efx
, drv
, efx
->net_dev
, "shutting down I/O\n");
1095 iounmap(efx
->membase
);
1096 efx
->membase
= NULL
;
1099 if (efx
->membase_phys
) {
1100 pci_release_region(efx
->pci_dev
, EFX_MEM_BAR
);
1101 efx
->membase_phys
= 0;
1104 pci_disable_device(efx
->pci_dev
);
1107 static int efx_wanted_parallelism(void)
1109 cpumask_var_t thread_mask
;
1116 if (unlikely(!zalloc_cpumask_var(&thread_mask
, GFP_KERNEL
))) {
1118 "sfc: RSS disabled due to allocation failure\n");
1123 for_each_online_cpu(cpu
) {
1124 if (!cpumask_test_cpu(cpu
, thread_mask
)) {
1126 cpumask_or(thread_mask
, thread_mask
,
1127 topology_thread_cpumask(cpu
));
1131 free_cpumask_var(thread_mask
);
1136 efx_init_rx_cpu_rmap(struct efx_nic
*efx
, struct msix_entry
*xentries
)
1138 #ifdef CONFIG_RFS_ACCEL
1141 efx
->net_dev
->rx_cpu_rmap
= alloc_irq_cpu_rmap(efx
->n_rx_channels
);
1142 if (!efx
->net_dev
->rx_cpu_rmap
)
1144 for (i
= 0; i
< efx
->n_rx_channels
; i
++) {
1145 rc
= irq_cpu_rmap_add(efx
->net_dev
->rx_cpu_rmap
,
1146 xentries
[i
].vector
);
1148 free_irq_cpu_rmap(efx
->net_dev
->rx_cpu_rmap
);
1149 efx
->net_dev
->rx_cpu_rmap
= NULL
;
1157 /* Probe the number and type of interrupts we are able to obtain, and
1158 * the resulting numbers of channels and RX queues.
1160 static int efx_probe_interrupts(struct efx_nic
*efx
)
1163 min_t(int, efx
->type
->phys_addr_channels
, EFX_MAX_CHANNELS
);
1166 if (efx
->interrupt_mode
== EFX_INT_MODE_MSIX
) {
1167 struct msix_entry xentries
[EFX_MAX_CHANNELS
];
1170 n_channels
= efx_wanted_parallelism();
1171 if (separate_tx_channels
)
1173 n_channels
= min(n_channels
, max_channels
);
1175 for (i
= 0; i
< n_channels
; i
++)
1176 xentries
[i
].entry
= i
;
1177 rc
= pci_enable_msix(efx
->pci_dev
, xentries
, n_channels
);
1179 netif_err(efx
, drv
, efx
->net_dev
,
1180 "WARNING: Insufficient MSI-X vectors"
1181 " available (%d < %d).\n", rc
, n_channels
);
1182 netif_err(efx
, drv
, efx
->net_dev
,
1183 "WARNING: Performance may be reduced.\n");
1184 EFX_BUG_ON_PARANOID(rc
>= n_channels
);
1186 rc
= pci_enable_msix(efx
->pci_dev
, xentries
,
1191 efx
->n_channels
= n_channels
;
1192 if (separate_tx_channels
) {
1193 efx
->n_tx_channels
=
1194 max(efx
->n_channels
/ 2, 1U);
1195 efx
->n_rx_channels
=
1196 max(efx
->n_channels
-
1197 efx
->n_tx_channels
, 1U);
1199 efx
->n_tx_channels
= efx
->n_channels
;
1200 efx
->n_rx_channels
= efx
->n_channels
;
1202 rc
= efx_init_rx_cpu_rmap(efx
, xentries
);
1204 pci_disable_msix(efx
->pci_dev
);
1207 for (i
= 0; i
< n_channels
; i
++)
1208 efx_get_channel(efx
, i
)->irq
=
1211 /* Fall back to single channel MSI */
1212 efx
->interrupt_mode
= EFX_INT_MODE_MSI
;
1213 netif_err(efx
, drv
, efx
->net_dev
,
1214 "could not enable MSI-X\n");
1218 /* Try single interrupt MSI */
1219 if (efx
->interrupt_mode
== EFX_INT_MODE_MSI
) {
1220 efx
->n_channels
= 1;
1221 efx
->n_rx_channels
= 1;
1222 efx
->n_tx_channels
= 1;
1223 rc
= pci_enable_msi(efx
->pci_dev
);
1225 efx_get_channel(efx
, 0)->irq
= efx
->pci_dev
->irq
;
1227 netif_err(efx
, drv
, efx
->net_dev
,
1228 "could not enable MSI\n");
1229 efx
->interrupt_mode
= EFX_INT_MODE_LEGACY
;
1233 /* Assume legacy interrupts */
1234 if (efx
->interrupt_mode
== EFX_INT_MODE_LEGACY
) {
1235 efx
->n_channels
= 1 + (separate_tx_channels
? 1 : 0);
1236 efx
->n_rx_channels
= 1;
1237 efx
->n_tx_channels
= 1;
1238 efx
->legacy_irq
= efx
->pci_dev
->irq
;
1244 /* Enable interrupts, then probe and start the event queues */
1245 static void efx_start_interrupts(struct efx_nic
*efx
)
1247 struct efx_channel
*channel
;
1249 if (efx
->legacy_irq
)
1250 efx
->legacy_irq_enabled
= true;
1251 efx_nic_enable_interrupts(efx
);
1253 efx_for_each_channel(channel
, efx
) {
1254 efx_init_eventq(channel
);
1255 efx_start_eventq(channel
);
1258 efx_mcdi_mode_event(efx
);
1261 static void efx_stop_interrupts(struct efx_nic
*efx
)
1263 struct efx_channel
*channel
;
1265 efx_mcdi_mode_poll(efx
);
1267 efx_nic_disable_interrupts(efx
);
1268 if (efx
->legacy_irq
) {
1269 synchronize_irq(efx
->legacy_irq
);
1270 efx
->legacy_irq_enabled
= false;
1273 efx_for_each_channel(channel
, efx
) {
1275 synchronize_irq(channel
->irq
);
1277 efx_stop_eventq(channel
);
1278 efx_fini_eventq(channel
);
1282 static void efx_remove_interrupts(struct efx_nic
*efx
)
1284 struct efx_channel
*channel
;
1286 /* Remove MSI/MSI-X interrupts */
1287 efx_for_each_channel(channel
, efx
)
1289 pci_disable_msi(efx
->pci_dev
);
1290 pci_disable_msix(efx
->pci_dev
);
1292 /* Remove legacy interrupt */
1293 efx
->legacy_irq
= 0;
1296 static void efx_set_channels(struct efx_nic
*efx
)
1298 struct efx_channel
*channel
;
1299 struct efx_tx_queue
*tx_queue
;
1301 efx
->tx_channel_offset
=
1302 separate_tx_channels
? efx
->n_channels
- efx
->n_tx_channels
: 0;
1304 /* We need to adjust the TX queue numbers if we have separate
1305 * RX-only and TX-only channels.
1307 efx_for_each_channel(channel
, efx
) {
1308 efx_for_each_channel_tx_queue(tx_queue
, channel
)
1309 tx_queue
->queue
-= (efx
->tx_channel_offset
*
1314 static int efx_probe_nic(struct efx_nic
*efx
)
1319 netif_dbg(efx
, probe
, efx
->net_dev
, "creating NIC\n");
1321 /* Carry out hardware-type specific initialisation */
1322 rc
= efx
->type
->probe(efx
);
1326 /* Determine the number of channels and queues by trying to hook
1327 * in MSI-X interrupts. */
1328 rc
= efx_probe_interrupts(efx
);
1332 if (efx
->n_channels
> 1)
1333 get_random_bytes(&efx
->rx_hash_key
, sizeof(efx
->rx_hash_key
));
1334 for (i
= 0; i
< ARRAY_SIZE(efx
->rx_indir_table
); i
++)
1335 efx
->rx_indir_table
[i
] =
1336 ethtool_rxfh_indir_default(i
, efx
->n_rx_channels
);
1338 efx_set_channels(efx
);
1339 netif_set_real_num_tx_queues(efx
->net_dev
, efx
->n_tx_channels
);
1340 netif_set_real_num_rx_queues(efx
->net_dev
, efx
->n_rx_channels
);
1342 /* Initialise the interrupt moderation settings */
1343 efx_init_irq_moderation(efx
, tx_irq_mod_usec
, rx_irq_mod_usec
, true,
1349 efx
->type
->remove(efx
);
1353 static void efx_remove_nic(struct efx_nic
*efx
)
1355 netif_dbg(efx
, drv
, efx
->net_dev
, "destroying NIC\n");
1357 efx_remove_interrupts(efx
);
1358 efx
->type
->remove(efx
);
1361 /**************************************************************************
1363 * NIC startup/shutdown
1365 *************************************************************************/
1367 static int efx_probe_all(struct efx_nic
*efx
)
1371 rc
= efx_probe_nic(efx
);
1373 netif_err(efx
, probe
, efx
->net_dev
, "failed to create NIC\n");
1377 rc
= efx_probe_port(efx
);
1379 netif_err(efx
, probe
, efx
->net_dev
, "failed to create port\n");
1383 efx
->rxq_entries
= efx
->txq_entries
= EFX_DEFAULT_DMAQ_SIZE
;
1384 rc
= efx_probe_channels(efx
);
1388 rc
= efx_probe_filters(efx
);
1390 netif_err(efx
, probe
, efx
->net_dev
,
1391 "failed to create filter tables\n");
1398 efx_remove_channels(efx
);
1400 efx_remove_port(efx
);
1402 efx_remove_nic(efx
);
1407 /* Called after previous invocation(s) of efx_stop_all, restarts the port,
1408 * kernel transmit queues and NAPI processing, and ensures that the port is
1409 * scheduled to be reconfigured. This function is safe to call multiple
1410 * times when the NIC is in any state.
1412 static void efx_start_all(struct efx_nic
*efx
)
1414 EFX_ASSERT_RESET_SERIALISED(efx
);
1416 /* Check that it is appropriate to restart the interface. All
1417 * of these flags are safe to read under just the rtnl lock */
1418 if (efx
->port_enabled
)
1420 if ((efx
->state
!= STATE_RUNNING
) && (efx
->state
!= STATE_INIT
))
1422 if (!netif_running(efx
->net_dev
))
1425 efx_start_port(efx
);
1426 efx_start_datapath(efx
);
1428 /* Start the hardware monitor if there is one. Otherwise (we're link
1429 * event driven), we have to poll the PHY because after an event queue
1430 * flush, we could have a missed a link state change */
1431 if (efx
->type
->monitor
!= NULL
) {
1432 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1433 efx_monitor_interval
);
1435 mutex_lock(&efx
->mac_lock
);
1436 if (efx
->phy_op
->poll(efx
))
1437 efx_link_status_changed(efx
);
1438 mutex_unlock(&efx
->mac_lock
);
1441 efx
->type
->start_stats(efx
);
1444 /* Flush all delayed work. Should only be called when no more delayed work
1445 * will be scheduled. This doesn't flush pending online resets (efx_reset),
1446 * since we're holding the rtnl_lock at this point. */
1447 static void efx_flush_all(struct efx_nic
*efx
)
1449 /* Make sure the hardware monitor is stopped */
1450 cancel_delayed_work_sync(&efx
->monitor_work
);
1451 /* Stop scheduled port reconfigurations */
1452 cancel_work_sync(&efx
->mac_work
);
1455 /* Quiesce hardware and software without bringing the link down.
1456 * Safe to call multiple times, when the nic and interface is in any
1457 * state. The caller is guaranteed to subsequently be in a position
1458 * to modify any hardware and software state they see fit without
1460 static void efx_stop_all(struct efx_nic
*efx
)
1462 EFX_ASSERT_RESET_SERIALISED(efx
);
1464 /* port_enabled can be read safely under the rtnl lock */
1465 if (!efx
->port_enabled
)
1468 efx
->type
->stop_stats(efx
);
1471 /* Flush efx_mac_work(), refill_workqueue, monitor_work */
1474 /* Stop the kernel transmit interface late, so the watchdog
1475 * timer isn't ticking over the flush */
1476 netif_tx_disable(efx
->net_dev
);
1478 efx_stop_datapath(efx
);
1481 static void efx_remove_all(struct efx_nic
*efx
)
1483 efx_remove_filters(efx
);
1484 efx_remove_channels(efx
);
1485 efx_remove_port(efx
);
1486 efx_remove_nic(efx
);
1489 /**************************************************************************
1491 * Interrupt moderation
1493 **************************************************************************/
1495 static unsigned int irq_mod_ticks(unsigned int usecs
, unsigned int quantum_ns
)
1499 if (usecs
* 1000 < quantum_ns
)
1500 return 1; /* never round down to 0 */
1501 return usecs
* 1000 / quantum_ns
;
1504 /* Set interrupt moderation parameters */
1505 int efx_init_irq_moderation(struct efx_nic
*efx
, unsigned int tx_usecs
,
1506 unsigned int rx_usecs
, bool rx_adaptive
,
1507 bool rx_may_override_tx
)
1509 struct efx_channel
*channel
;
1510 unsigned int irq_mod_max
= DIV_ROUND_UP(efx
->type
->timer_period_max
*
1511 efx
->timer_quantum_ns
,
1513 unsigned int tx_ticks
;
1514 unsigned int rx_ticks
;
1516 EFX_ASSERT_RESET_SERIALISED(efx
);
1518 if (tx_usecs
> irq_mod_max
|| rx_usecs
> irq_mod_max
)
1521 tx_ticks
= irq_mod_ticks(tx_usecs
, efx
->timer_quantum_ns
);
1522 rx_ticks
= irq_mod_ticks(rx_usecs
, efx
->timer_quantum_ns
);
1524 if (tx_ticks
!= rx_ticks
&& efx
->tx_channel_offset
== 0 &&
1525 !rx_may_override_tx
) {
1526 netif_err(efx
, drv
, efx
->net_dev
, "Channels are shared. "
1527 "RX and TX IRQ moderation must be equal\n");
1531 efx
->irq_rx_adaptive
= rx_adaptive
;
1532 efx
->irq_rx_moderation
= rx_ticks
;
1533 efx_for_each_channel(channel
, efx
) {
1534 if (efx_channel_has_rx_queue(channel
))
1535 channel
->irq_moderation
= rx_ticks
;
1536 else if (efx_channel_has_tx_queues(channel
))
1537 channel
->irq_moderation
= tx_ticks
;
1543 void efx_get_irq_moderation(struct efx_nic
*efx
, unsigned int *tx_usecs
,
1544 unsigned int *rx_usecs
, bool *rx_adaptive
)
1546 /* We must round up when converting ticks to microseconds
1547 * because we round down when converting the other way.
1550 *rx_adaptive
= efx
->irq_rx_adaptive
;
1551 *rx_usecs
= DIV_ROUND_UP(efx
->irq_rx_moderation
*
1552 efx
->timer_quantum_ns
,
1555 /* If channels are shared between RX and TX, so is IRQ
1556 * moderation. Otherwise, IRQ moderation is the same for all
1557 * TX channels and is not adaptive.
1559 if (efx
->tx_channel_offset
== 0)
1560 *tx_usecs
= *rx_usecs
;
1562 *tx_usecs
= DIV_ROUND_UP(
1563 efx
->channel
[efx
->tx_channel_offset
]->irq_moderation
*
1564 efx
->timer_quantum_ns
,
1568 /**************************************************************************
1572 **************************************************************************/
1574 /* Run periodically off the general workqueue */
1575 static void efx_monitor(struct work_struct
*data
)
1577 struct efx_nic
*efx
= container_of(data
, struct efx_nic
,
1580 netif_vdbg(efx
, timer
, efx
->net_dev
,
1581 "hardware monitor executing on CPU %d\n",
1582 raw_smp_processor_id());
1583 BUG_ON(efx
->type
->monitor
== NULL
);
1585 /* If the mac_lock is already held then it is likely a port
1586 * reconfiguration is already in place, which will likely do
1587 * most of the work of monitor() anyway. */
1588 if (mutex_trylock(&efx
->mac_lock
)) {
1589 if (efx
->port_enabled
)
1590 efx
->type
->monitor(efx
);
1591 mutex_unlock(&efx
->mac_lock
);
1594 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1595 efx_monitor_interval
);
1598 /**************************************************************************
1602 *************************************************************************/
1605 * Context: process, rtnl_lock() held.
1607 static int efx_ioctl(struct net_device
*net_dev
, struct ifreq
*ifr
, int cmd
)
1609 struct efx_nic
*efx
= netdev_priv(net_dev
);
1610 struct mii_ioctl_data
*data
= if_mii(ifr
);
1612 EFX_ASSERT_RESET_SERIALISED(efx
);
1614 /* Convert phy_id from older PRTAD/DEVAD format */
1615 if ((cmd
== SIOCGMIIREG
|| cmd
== SIOCSMIIREG
) &&
1616 (data
->phy_id
& 0xfc00) == 0x0400)
1617 data
->phy_id
^= MDIO_PHY_ID_C45
| 0x0400;
1619 return mdio_mii_ioctl(&efx
->mdio
, data
, cmd
);
1622 /**************************************************************************
1626 **************************************************************************/
1628 static void efx_init_napi(struct efx_nic
*efx
)
1630 struct efx_channel
*channel
;
1632 efx_for_each_channel(channel
, efx
) {
1633 channel
->napi_dev
= efx
->net_dev
;
1634 netif_napi_add(channel
->napi_dev
, &channel
->napi_str
,
1635 efx_poll
, napi_weight
);
1639 static void efx_fini_napi_channel(struct efx_channel
*channel
)
1641 if (channel
->napi_dev
)
1642 netif_napi_del(&channel
->napi_str
);
1643 channel
->napi_dev
= NULL
;
1646 static void efx_fini_napi(struct efx_nic
*efx
)
1648 struct efx_channel
*channel
;
1650 efx_for_each_channel(channel
, efx
)
1651 efx_fini_napi_channel(channel
);
1654 /**************************************************************************
1656 * Kernel netpoll interface
1658 *************************************************************************/
1660 #ifdef CONFIG_NET_POLL_CONTROLLER
1662 /* Although in the common case interrupts will be disabled, this is not
1663 * guaranteed. However, all our work happens inside the NAPI callback,
1664 * so no locking is required.
1666 static void efx_netpoll(struct net_device
*net_dev
)
1668 struct efx_nic
*efx
= netdev_priv(net_dev
);
1669 struct efx_channel
*channel
;
1671 efx_for_each_channel(channel
, efx
)
1672 efx_schedule_channel(channel
);
1677 /**************************************************************************
1679 * Kernel net device interface
1681 *************************************************************************/
1683 /* Context: process, rtnl_lock() held. */
1684 static int efx_net_open(struct net_device
*net_dev
)
1686 struct efx_nic
*efx
= netdev_priv(net_dev
);
1687 EFX_ASSERT_RESET_SERIALISED(efx
);
1689 netif_dbg(efx
, ifup
, efx
->net_dev
, "opening device on CPU %d\n",
1690 raw_smp_processor_id());
1692 if (efx
->state
== STATE_DISABLED
)
1694 if (efx
->phy_mode
& PHY_MODE_SPECIAL
)
1696 if (efx_mcdi_poll_reboot(efx
) && efx_reset(efx
, RESET_TYPE_ALL
))
1699 /* Notify the kernel of the link state polled during driver load,
1700 * before the monitor starts running */
1701 efx_link_status_changed(efx
);
1707 /* Context: process, rtnl_lock() held.
1708 * Note that the kernel will ignore our return code; this method
1709 * should really be a void.
1711 static int efx_net_stop(struct net_device
*net_dev
)
1713 struct efx_nic
*efx
= netdev_priv(net_dev
);
1715 netif_dbg(efx
, ifdown
, efx
->net_dev
, "closing on CPU %d\n",
1716 raw_smp_processor_id());
1718 if (efx
->state
!= STATE_DISABLED
) {
1719 /* Stop the device and flush all the channels */
1726 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1727 static struct rtnl_link_stats64
*efx_net_stats(struct net_device
*net_dev
,
1728 struct rtnl_link_stats64
*stats
)
1730 struct efx_nic
*efx
= netdev_priv(net_dev
);
1731 struct efx_mac_stats
*mac_stats
= &efx
->mac_stats
;
1733 spin_lock_bh(&efx
->stats_lock
);
1735 efx
->type
->update_stats(efx
);
1737 stats
->rx_packets
= mac_stats
->rx_packets
;
1738 stats
->tx_packets
= mac_stats
->tx_packets
;
1739 stats
->rx_bytes
= mac_stats
->rx_bytes
;
1740 stats
->tx_bytes
= mac_stats
->tx_bytes
;
1741 stats
->rx_dropped
= efx
->n_rx_nodesc_drop_cnt
;
1742 stats
->multicast
= mac_stats
->rx_multicast
;
1743 stats
->collisions
= mac_stats
->tx_collision
;
1744 stats
->rx_length_errors
= (mac_stats
->rx_gtjumbo
+
1745 mac_stats
->rx_length_error
);
1746 stats
->rx_crc_errors
= mac_stats
->rx_bad
;
1747 stats
->rx_frame_errors
= mac_stats
->rx_align_error
;
1748 stats
->rx_fifo_errors
= mac_stats
->rx_overflow
;
1749 stats
->rx_missed_errors
= mac_stats
->rx_missed
;
1750 stats
->tx_window_errors
= mac_stats
->tx_late_collision
;
1752 stats
->rx_errors
= (stats
->rx_length_errors
+
1753 stats
->rx_crc_errors
+
1754 stats
->rx_frame_errors
+
1755 mac_stats
->rx_symbol_error
);
1756 stats
->tx_errors
= (stats
->tx_window_errors
+
1759 spin_unlock_bh(&efx
->stats_lock
);
1764 /* Context: netif_tx_lock held, BHs disabled. */
1765 static void efx_watchdog(struct net_device
*net_dev
)
1767 struct efx_nic
*efx
= netdev_priv(net_dev
);
1769 netif_err(efx
, tx_err
, efx
->net_dev
,
1770 "TX stuck with port_enabled=%d: resetting channels\n",
1773 efx_schedule_reset(efx
, RESET_TYPE_TX_WATCHDOG
);
1777 /* Context: process, rtnl_lock() held. */
1778 static int efx_change_mtu(struct net_device
*net_dev
, int new_mtu
)
1780 struct efx_nic
*efx
= netdev_priv(net_dev
);
1782 EFX_ASSERT_RESET_SERIALISED(efx
);
1784 if (new_mtu
> EFX_MAX_MTU
)
1789 netif_dbg(efx
, drv
, efx
->net_dev
, "changing MTU to %d\n", new_mtu
);
1791 mutex_lock(&efx
->mac_lock
);
1792 /* Reconfigure the MAC before enabling the dma queues so that
1793 * the RX buffers don't overflow */
1794 net_dev
->mtu
= new_mtu
;
1795 efx
->type
->reconfigure_mac(efx
);
1796 mutex_unlock(&efx
->mac_lock
);
1802 static int efx_set_mac_address(struct net_device
*net_dev
, void *data
)
1804 struct efx_nic
*efx
= netdev_priv(net_dev
);
1805 struct sockaddr
*addr
= data
;
1806 char *new_addr
= addr
->sa_data
;
1808 EFX_ASSERT_RESET_SERIALISED(efx
);
1810 if (!is_valid_ether_addr(new_addr
)) {
1811 netif_err(efx
, drv
, efx
->net_dev
,
1812 "invalid ethernet MAC address requested: %pM\n",
1817 memcpy(net_dev
->dev_addr
, new_addr
, net_dev
->addr_len
);
1819 /* Reconfigure the MAC */
1820 mutex_lock(&efx
->mac_lock
);
1821 efx
->type
->reconfigure_mac(efx
);
1822 mutex_unlock(&efx
->mac_lock
);
1827 /* Context: netif_addr_lock held, BHs disabled. */
1828 static void efx_set_rx_mode(struct net_device
*net_dev
)
1830 struct efx_nic
*efx
= netdev_priv(net_dev
);
1831 struct netdev_hw_addr
*ha
;
1832 union efx_multicast_hash
*mc_hash
= &efx
->multicast_hash
;
1836 efx
->promiscuous
= !!(net_dev
->flags
& IFF_PROMISC
);
1838 /* Build multicast hash table */
1839 if (efx
->promiscuous
|| (net_dev
->flags
& IFF_ALLMULTI
)) {
1840 memset(mc_hash
, 0xff, sizeof(*mc_hash
));
1842 memset(mc_hash
, 0x00, sizeof(*mc_hash
));
1843 netdev_for_each_mc_addr(ha
, net_dev
) {
1844 crc
= ether_crc_le(ETH_ALEN
, ha
->addr
);
1845 bit
= crc
& (EFX_MCAST_HASH_ENTRIES
- 1);
1846 set_bit_le(bit
, mc_hash
->byte
);
1849 /* Broadcast packets go through the multicast hash filter.
1850 * ether_crc_le() of the broadcast address is 0xbe2612ff
1851 * so we always add bit 0xff to the mask.
1853 set_bit_le(0xff, mc_hash
->byte
);
1856 if (efx
->port_enabled
)
1857 queue_work(efx
->workqueue
, &efx
->mac_work
);
1858 /* Otherwise efx_start_port() will do this */
1861 static int efx_set_features(struct net_device
*net_dev
, netdev_features_t data
)
1863 struct efx_nic
*efx
= netdev_priv(net_dev
);
1865 /* If disabling RX n-tuple filtering, clear existing filters */
1866 if (net_dev
->features
& ~data
& NETIF_F_NTUPLE
)
1867 efx_filter_clear_rx(efx
, EFX_FILTER_PRI_MANUAL
);
1872 static const struct net_device_ops efx_netdev_ops
= {
1873 .ndo_open
= efx_net_open
,
1874 .ndo_stop
= efx_net_stop
,
1875 .ndo_get_stats64
= efx_net_stats
,
1876 .ndo_tx_timeout
= efx_watchdog
,
1877 .ndo_start_xmit
= efx_hard_start_xmit
,
1878 .ndo_validate_addr
= eth_validate_addr
,
1879 .ndo_do_ioctl
= efx_ioctl
,
1880 .ndo_change_mtu
= efx_change_mtu
,
1881 .ndo_set_mac_address
= efx_set_mac_address
,
1882 .ndo_set_rx_mode
= efx_set_rx_mode
,
1883 .ndo_set_features
= efx_set_features
,
1884 #ifdef CONFIG_NET_POLL_CONTROLLER
1885 .ndo_poll_controller
= efx_netpoll
,
1887 .ndo_setup_tc
= efx_setup_tc
,
1888 #ifdef CONFIG_RFS_ACCEL
1889 .ndo_rx_flow_steer
= efx_filter_rfs
,
1893 static void efx_update_name(struct efx_nic
*efx
)
1895 strcpy(efx
->name
, efx
->net_dev
->name
);
1896 efx_mtd_rename(efx
);
1897 efx_set_channel_names(efx
);
1900 static int efx_netdev_event(struct notifier_block
*this,
1901 unsigned long event
, void *ptr
)
1903 struct net_device
*net_dev
= ptr
;
1905 if (net_dev
->netdev_ops
== &efx_netdev_ops
&&
1906 event
== NETDEV_CHANGENAME
)
1907 efx_update_name(netdev_priv(net_dev
));
1912 static struct notifier_block efx_netdev_notifier
= {
1913 .notifier_call
= efx_netdev_event
,
1917 show_phy_type(struct device
*dev
, struct device_attribute
*attr
, char *buf
)
1919 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
1920 return sprintf(buf
, "%d\n", efx
->phy_type
);
1922 static DEVICE_ATTR(phy_type
, 0644, show_phy_type
, NULL
);
1924 static int efx_register_netdev(struct efx_nic
*efx
)
1926 struct net_device
*net_dev
= efx
->net_dev
;
1927 struct efx_channel
*channel
;
1930 net_dev
->watchdog_timeo
= 5 * HZ
;
1931 net_dev
->irq
= efx
->pci_dev
->irq
;
1932 net_dev
->netdev_ops
= &efx_netdev_ops
;
1933 SET_ETHTOOL_OPS(net_dev
, &efx_ethtool_ops
);
1937 rc
= dev_alloc_name(net_dev
, net_dev
->name
);
1940 efx_update_name(efx
);
1942 rc
= register_netdevice(net_dev
);
1946 efx_for_each_channel(channel
, efx
) {
1947 struct efx_tx_queue
*tx_queue
;
1948 efx_for_each_channel_tx_queue(tx_queue
, channel
)
1949 efx_init_tx_queue_core_txq(tx_queue
);
1952 /* Always start with carrier off; PHY events will detect the link */
1953 netif_carrier_off(net_dev
);
1957 rc
= device_create_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
1959 netif_err(efx
, drv
, efx
->net_dev
,
1960 "failed to init net dev attributes\n");
1961 goto fail_registered
;
1968 netif_err(efx
, drv
, efx
->net_dev
, "could not register net dev\n");
1972 unregister_netdev(net_dev
);
1976 static void efx_unregister_netdev(struct efx_nic
*efx
)
1978 struct efx_channel
*channel
;
1979 struct efx_tx_queue
*tx_queue
;
1984 BUG_ON(netdev_priv(efx
->net_dev
) != efx
);
1986 /* Free up any skbs still remaining. This has to happen before
1987 * we try to unregister the netdev as running their destructors
1988 * may be needed to get the device ref. count to 0. */
1989 efx_for_each_channel(channel
, efx
) {
1990 efx_for_each_channel_tx_queue(tx_queue
, channel
)
1991 efx_release_tx_buffers(tx_queue
);
1994 strlcpy(efx
->name
, pci_name(efx
->pci_dev
), sizeof(efx
->name
));
1995 device_remove_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
1996 unregister_netdev(efx
->net_dev
);
1999 /**************************************************************************
2001 * Device reset and suspend
2003 **************************************************************************/
2005 /* Tears down the entire software state and most of the hardware state
2007 void efx_reset_down(struct efx_nic
*efx
, enum reset_type method
)
2009 EFX_ASSERT_RESET_SERIALISED(efx
);
2012 mutex_lock(&efx
->mac_lock
);
2014 efx_stop_interrupts(efx
);
2015 if (efx
->port_initialized
&& method
!= RESET_TYPE_INVISIBLE
)
2016 efx
->phy_op
->fini(efx
);
2017 efx
->type
->fini(efx
);
2020 /* This function will always ensure that the locks acquired in
2021 * efx_reset_down() are released. A failure return code indicates
2022 * that we were unable to reinitialise the hardware, and the
2023 * driver should be disabled. If ok is false, then the rx and tx
2024 * engines are not restarted, pending a RESET_DISABLE. */
2025 int efx_reset_up(struct efx_nic
*efx
, enum reset_type method
, bool ok
)
2029 EFX_ASSERT_RESET_SERIALISED(efx
);
2031 rc
= efx
->type
->init(efx
);
2033 netif_err(efx
, drv
, efx
->net_dev
, "failed to initialise NIC\n");
2040 if (efx
->port_initialized
&& method
!= RESET_TYPE_INVISIBLE
) {
2041 rc
= efx
->phy_op
->init(efx
);
2044 if (efx
->phy_op
->reconfigure(efx
))
2045 netif_err(efx
, drv
, efx
->net_dev
,
2046 "could not restore PHY settings\n");
2049 efx
->type
->reconfigure_mac(efx
);
2051 efx_start_interrupts(efx
);
2052 efx_restore_filters(efx
);
2054 mutex_unlock(&efx
->mac_lock
);
2061 efx
->port_initialized
= false;
2063 mutex_unlock(&efx
->mac_lock
);
2068 /* Reset the NIC using the specified method. Note that the reset may
2069 * fail, in which case the card will be left in an unusable state.
2071 * Caller must hold the rtnl_lock.
2073 int efx_reset(struct efx_nic
*efx
, enum reset_type method
)
2078 netif_info(efx
, drv
, efx
->net_dev
, "resetting (%s)\n",
2079 RESET_TYPE(method
));
2081 netif_device_detach(efx
->net_dev
);
2082 efx_reset_down(efx
, method
);
2084 rc
= efx
->type
->reset(efx
, method
);
2086 netif_err(efx
, drv
, efx
->net_dev
, "failed to reset hardware\n");
2090 /* Clear flags for the scopes we covered. We assume the NIC and
2091 * driver are now quiescent so that there is no race here.
2093 efx
->reset_pending
&= -(1 << (method
+ 1));
2095 /* Reinitialise bus-mastering, which may have been turned off before
2096 * the reset was scheduled. This is still appropriate, even in the
2097 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2098 * can respond to requests. */
2099 pci_set_master(efx
->pci_dev
);
2102 /* Leave device stopped if necessary */
2103 disabled
= rc
|| method
== RESET_TYPE_DISABLE
;
2104 rc2
= efx_reset_up(efx
, method
, !disabled
);
2112 dev_close(efx
->net_dev
);
2113 netif_err(efx
, drv
, efx
->net_dev
, "has been disabled\n");
2114 efx
->state
= STATE_DISABLED
;
2116 netif_dbg(efx
, drv
, efx
->net_dev
, "reset complete\n");
2117 netif_device_attach(efx
->net_dev
);
2122 /* The worker thread exists so that code that cannot sleep can
2123 * schedule a reset for later.
2125 static void efx_reset_work(struct work_struct
*data
)
2127 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, reset_work
);
2128 unsigned long pending
= ACCESS_ONCE(efx
->reset_pending
);
2133 /* If we're not RUNNING then don't reset. Leave the reset_pending
2134 * flags set so that efx_pci_probe_main will be retried */
2135 if (efx
->state
!= STATE_RUNNING
) {
2136 netif_info(efx
, drv
, efx
->net_dev
,
2137 "scheduled reset quenched. NIC not RUNNING\n");
2142 (void)efx_reset(efx
, fls(pending
) - 1);
2146 void efx_schedule_reset(struct efx_nic
*efx
, enum reset_type type
)
2148 enum reset_type method
;
2151 case RESET_TYPE_INVISIBLE
:
2152 case RESET_TYPE_ALL
:
2153 case RESET_TYPE_WORLD
:
2154 case RESET_TYPE_DISABLE
:
2156 netif_dbg(efx
, drv
, efx
->net_dev
, "scheduling %s reset\n",
2157 RESET_TYPE(method
));
2160 method
= efx
->type
->map_reset_reason(type
);
2161 netif_dbg(efx
, drv
, efx
->net_dev
,
2162 "scheduling %s reset for %s\n",
2163 RESET_TYPE(method
), RESET_TYPE(type
));
2167 set_bit(method
, &efx
->reset_pending
);
2169 /* efx_process_channel() will no longer read events once a
2170 * reset is scheduled. So switch back to poll'd MCDI completions. */
2171 efx_mcdi_mode_poll(efx
);
2173 queue_work(reset_workqueue
, &efx
->reset_work
);
2176 /**************************************************************************
2178 * List of NICs we support
2180 **************************************************************************/
2182 /* PCI device ID table */
2183 static DEFINE_PCI_DEVICE_TABLE(efx_pci_table
) = {
2184 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
,
2185 PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0
),
2186 .driver_data
= (unsigned long) &falcon_a1_nic_type
},
2187 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
,
2188 PCI_DEVICE_ID_SOLARFLARE_SFC4000B
),
2189 .driver_data
= (unsigned long) &falcon_b0_nic_type
},
2190 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
, 0x0803), /* SFC9020 */
2191 .driver_data
= (unsigned long) &siena_a0_nic_type
},
2192 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
, 0x0813), /* SFL9021 */
2193 .driver_data
= (unsigned long) &siena_a0_nic_type
},
2194 {0} /* end of list */
2197 /**************************************************************************
2199 * Dummy PHY/MAC operations
2201 * Can be used for some unimplemented operations
2202 * Needed so all function pointers are valid and do not have to be tested
2205 **************************************************************************/
2206 int efx_port_dummy_op_int(struct efx_nic
*efx
)
2210 void efx_port_dummy_op_void(struct efx_nic
*efx
) {}
2212 static bool efx_port_dummy_op_poll(struct efx_nic
*efx
)
2217 static const struct efx_phy_operations efx_dummy_phy_operations
= {
2218 .init
= efx_port_dummy_op_int
,
2219 .reconfigure
= efx_port_dummy_op_int
,
2220 .poll
= efx_port_dummy_op_poll
,
2221 .fini
= efx_port_dummy_op_void
,
2224 /**************************************************************************
2228 **************************************************************************/
2230 /* This zeroes out and then fills in the invariants in a struct
2231 * efx_nic (including all sub-structures).
2233 static int efx_init_struct(struct efx_nic
*efx
, const struct efx_nic_type
*type
,
2234 struct pci_dev
*pci_dev
, struct net_device
*net_dev
)
2238 /* Initialise common structures */
2239 memset(efx
, 0, sizeof(*efx
));
2240 spin_lock_init(&efx
->biu_lock
);
2241 #ifdef CONFIG_SFC_MTD
2242 INIT_LIST_HEAD(&efx
->mtd_list
);
2244 INIT_WORK(&efx
->reset_work
, efx_reset_work
);
2245 INIT_DELAYED_WORK(&efx
->monitor_work
, efx_monitor
);
2246 efx
->pci_dev
= pci_dev
;
2247 efx
->msg_enable
= debug
;
2248 efx
->state
= STATE_INIT
;
2249 strlcpy(efx
->name
, pci_name(pci_dev
), sizeof(efx
->name
));
2251 efx
->net_dev
= net_dev
;
2252 spin_lock_init(&efx
->stats_lock
);
2253 mutex_init(&efx
->mac_lock
);
2254 efx
->phy_op
= &efx_dummy_phy_operations
;
2255 efx
->mdio
.dev
= net_dev
;
2256 INIT_WORK(&efx
->mac_work
, efx_mac_work
);
2257 init_waitqueue_head(&efx
->flush_wq
);
2259 for (i
= 0; i
< EFX_MAX_CHANNELS
; i
++) {
2260 efx
->channel
[i
] = efx_alloc_channel(efx
, i
, NULL
);
2261 if (!efx
->channel
[i
])
2267 EFX_BUG_ON_PARANOID(efx
->type
->phys_addr_channels
> EFX_MAX_CHANNELS
);
2269 /* Higher numbered interrupt modes are less capable! */
2270 efx
->interrupt_mode
= max(efx
->type
->max_interrupt_mode
,
2273 /* Would be good to use the net_dev name, but we're too early */
2274 snprintf(efx
->workqueue_name
, sizeof(efx
->workqueue_name
), "sfc%s",
2276 efx
->workqueue
= create_singlethread_workqueue(efx
->workqueue_name
);
2277 if (!efx
->workqueue
)
2283 efx_fini_struct(efx
);
2287 static void efx_fini_struct(struct efx_nic
*efx
)
2291 for (i
= 0; i
< EFX_MAX_CHANNELS
; i
++)
2292 kfree(efx
->channel
[i
]);
2294 if (efx
->workqueue
) {
2295 destroy_workqueue(efx
->workqueue
);
2296 efx
->workqueue
= NULL
;
2300 /**************************************************************************
2304 **************************************************************************/
2306 /* Main body of final NIC shutdown code
2307 * This is called only at module unload (or hotplug removal).
2309 static void efx_pci_remove_main(struct efx_nic
*efx
)
2311 #ifdef CONFIG_RFS_ACCEL
2312 free_irq_cpu_rmap(efx
->net_dev
->rx_cpu_rmap
);
2313 efx
->net_dev
->rx_cpu_rmap
= NULL
;
2315 efx_stop_interrupts(efx
);
2316 efx_nic_fini_interrupt(efx
);
2318 efx
->type
->fini(efx
);
2320 efx_remove_all(efx
);
2323 /* Final NIC shutdown
2324 * This is called only at module unload (or hotplug removal).
2326 static void efx_pci_remove(struct pci_dev
*pci_dev
)
2328 struct efx_nic
*efx
;
2330 efx
= pci_get_drvdata(pci_dev
);
2334 /* Mark the NIC as fini, then stop the interface */
2336 efx
->state
= STATE_FINI
;
2337 dev_close(efx
->net_dev
);
2339 /* Allow any queued efx_resets() to complete */
2342 efx_stop_interrupts(efx
);
2343 efx_unregister_netdev(efx
);
2345 efx_mtd_remove(efx
);
2347 /* Wait for any scheduled resets to complete. No more will be
2348 * scheduled from this point because efx_stop_all() has been
2349 * called, we are no longer registered with driverlink, and
2350 * the net_device's have been removed. */
2351 cancel_work_sync(&efx
->reset_work
);
2353 efx_pci_remove_main(efx
);
2356 netif_dbg(efx
, drv
, efx
->net_dev
, "shutdown successful\n");
2358 pci_set_drvdata(pci_dev
, NULL
);
2359 efx_fini_struct(efx
);
2360 free_netdev(efx
->net_dev
);
2363 /* Main body of NIC initialisation
2364 * This is called at module load (or hotplug insertion, theoretically).
2366 static int efx_pci_probe_main(struct efx_nic
*efx
)
2370 /* Do start-of-day initialisation */
2371 rc
= efx_probe_all(efx
);
2377 rc
= efx
->type
->init(efx
);
2379 netif_err(efx
, probe
, efx
->net_dev
,
2380 "failed to initialise NIC\n");
2384 rc
= efx_init_port(efx
);
2386 netif_err(efx
, probe
, efx
->net_dev
,
2387 "failed to initialise port\n");
2391 rc
= efx_nic_init_interrupt(efx
);
2394 efx_start_interrupts(efx
);
2401 efx
->type
->fini(efx
);
2404 efx_remove_all(efx
);
2409 /* NIC initialisation
2411 * This is called at module load (or hotplug insertion,
2412 * theoretically). It sets up PCI mappings, resets the NIC,
2413 * sets up and registers the network devices with the kernel and hooks
2414 * the interrupt service routine. It does not prepare the device for
2415 * transmission; this is left to the first time one of the network
2416 * interfaces is brought up (i.e. efx_net_open).
2418 static int __devinit
efx_pci_probe(struct pci_dev
*pci_dev
,
2419 const struct pci_device_id
*entry
)
2421 const struct efx_nic_type
*type
= (const struct efx_nic_type
*) entry
->driver_data
;
2422 struct net_device
*net_dev
;
2423 struct efx_nic
*efx
;
2426 /* Allocate and initialise a struct net_device and struct efx_nic */
2427 net_dev
= alloc_etherdev_mqs(sizeof(*efx
), EFX_MAX_CORE_TX_QUEUES
,
2431 net_dev
->features
|= (type
->offload_features
| NETIF_F_SG
|
2432 NETIF_F_HIGHDMA
| NETIF_F_TSO
|
2434 if (type
->offload_features
& NETIF_F_V6_CSUM
)
2435 net_dev
->features
|= NETIF_F_TSO6
;
2436 /* Mask for features that also apply to VLAN devices */
2437 net_dev
->vlan_features
|= (NETIF_F_ALL_CSUM
| NETIF_F_SG
|
2438 NETIF_F_HIGHDMA
| NETIF_F_ALL_TSO
|
2440 /* All offloads can be toggled */
2441 net_dev
->hw_features
= net_dev
->features
& ~NETIF_F_HIGHDMA
;
2442 efx
= netdev_priv(net_dev
);
2443 pci_set_drvdata(pci_dev
, efx
);
2444 SET_NETDEV_DEV(net_dev
, &pci_dev
->dev
);
2445 rc
= efx_init_struct(efx
, type
, pci_dev
, net_dev
);
2449 netif_info(efx
, probe
, efx
->net_dev
,
2450 "Solarflare NIC detected\n");
2452 /* Set up basic I/O (BAR mappings etc) */
2453 rc
= efx_init_io(efx
);
2457 rc
= efx_pci_probe_main(efx
);
2459 /* Serialise against efx_reset(). No more resets will be
2460 * scheduled since efx_stop_all() has been called, and we have
2461 * not and never have been registered.
2463 cancel_work_sync(&efx
->reset_work
);
2468 /* If there was a scheduled reset during probe, the NIC is
2469 * probably hosed anyway.
2471 if (efx
->reset_pending
) {
2476 /* Switch to the running state before we expose the device to the OS,
2477 * so that dev_open()|efx_start_all() will actually start the device */
2478 efx
->state
= STATE_RUNNING
;
2480 rc
= efx_register_netdev(efx
);
2484 netif_dbg(efx
, probe
, efx
->net_dev
, "initialisation successful\n");
2486 /* Try to create MTDs, but allow this to fail */
2488 rc
= efx_mtd_probe(efx
);
2491 netif_warn(efx
, probe
, efx
->net_dev
,
2492 "failed to create MTDs (%d)\n", rc
);
2497 efx_pci_remove_main(efx
);
2501 efx_fini_struct(efx
);
2504 netif_dbg(efx
, drv
, efx
->net_dev
, "initialisation failed. rc=%d\n", rc
);
2505 free_netdev(net_dev
);
2509 static int efx_pm_freeze(struct device
*dev
)
2511 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
2513 efx
->state
= STATE_FINI
;
2515 netif_device_detach(efx
->net_dev
);
2518 efx_stop_interrupts(efx
);
2523 static int efx_pm_thaw(struct device
*dev
)
2525 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
2527 efx
->state
= STATE_INIT
;
2529 efx_start_interrupts(efx
);
2531 mutex_lock(&efx
->mac_lock
);
2532 efx
->phy_op
->reconfigure(efx
);
2533 mutex_unlock(&efx
->mac_lock
);
2537 netif_device_attach(efx
->net_dev
);
2539 efx
->state
= STATE_RUNNING
;
2541 efx
->type
->resume_wol(efx
);
2543 /* Reschedule any quenched resets scheduled during efx_pm_freeze() */
2544 queue_work(reset_workqueue
, &efx
->reset_work
);
2549 static int efx_pm_poweroff(struct device
*dev
)
2551 struct pci_dev
*pci_dev
= to_pci_dev(dev
);
2552 struct efx_nic
*efx
= pci_get_drvdata(pci_dev
);
2554 efx
->type
->fini(efx
);
2556 efx
->reset_pending
= 0;
2558 pci_save_state(pci_dev
);
2559 return pci_set_power_state(pci_dev
, PCI_D3hot
);
2562 /* Used for both resume and restore */
2563 static int efx_pm_resume(struct device
*dev
)
2565 struct pci_dev
*pci_dev
= to_pci_dev(dev
);
2566 struct efx_nic
*efx
= pci_get_drvdata(pci_dev
);
2569 rc
= pci_set_power_state(pci_dev
, PCI_D0
);
2572 pci_restore_state(pci_dev
);
2573 rc
= pci_enable_device(pci_dev
);
2576 pci_set_master(efx
->pci_dev
);
2577 rc
= efx
->type
->reset(efx
, RESET_TYPE_ALL
);
2580 rc
= efx
->type
->init(efx
);
2587 static int efx_pm_suspend(struct device
*dev
)
2592 rc
= efx_pm_poweroff(dev
);
2598 static const struct dev_pm_ops efx_pm_ops
= {
2599 .suspend
= efx_pm_suspend
,
2600 .resume
= efx_pm_resume
,
2601 .freeze
= efx_pm_freeze
,
2602 .thaw
= efx_pm_thaw
,
2603 .poweroff
= efx_pm_poweroff
,
2604 .restore
= efx_pm_resume
,
2607 static struct pci_driver efx_pci_driver
= {
2608 .name
= KBUILD_MODNAME
,
2609 .id_table
= efx_pci_table
,
2610 .probe
= efx_pci_probe
,
2611 .remove
= efx_pci_remove
,
2612 .driver
.pm
= &efx_pm_ops
,
2615 /**************************************************************************
2617 * Kernel module interface
2619 *************************************************************************/
2621 module_param(interrupt_mode
, uint
, 0444);
2622 MODULE_PARM_DESC(interrupt_mode
,
2623 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
2625 static int __init
efx_init_module(void)
2629 printk(KERN_INFO
"Solarflare NET driver v" EFX_DRIVER_VERSION
"\n");
2631 rc
= register_netdevice_notifier(&efx_netdev_notifier
);
2635 reset_workqueue
= create_singlethread_workqueue("sfc_reset");
2636 if (!reset_workqueue
) {
2641 rc
= pci_register_driver(&efx_pci_driver
);
2648 destroy_workqueue(reset_workqueue
);
2650 unregister_netdevice_notifier(&efx_netdev_notifier
);
2655 static void __exit
efx_exit_module(void)
2657 printk(KERN_INFO
"Solarflare NET driver unloading\n");
2659 pci_unregister_driver(&efx_pci_driver
);
2660 destroy_workqueue(reset_workqueue
);
2661 unregister_netdevice_notifier(&efx_netdev_notifier
);
2665 module_init(efx_init_module
);
2666 module_exit(efx_exit_module
);
2668 MODULE_AUTHOR("Solarflare Communications and "
2669 "Michael Brown <mbrown@fensystems.co.uk>");
2670 MODULE_DESCRIPTION("Solarflare Communications network driver");
2671 MODULE_LICENSE("GPL");
2672 MODULE_DEVICE_TABLE(pci
, efx_pci_table
);